| The collection of optical remote sensing images is limited by factors such as equipment and environment.The amount of data is scarce,and the size of surface objects captured from high altitude varies significantly,including a large number of multiple categories of objects and complex backgrounds,making it difficult to extract key information for identifying image targets.At present,algorithms for natural image object detection with deep learning as the mainstream ideology are difficult to cope with complex and ever-changing remote sensing scenes.In order to fully explore the feature information of each remote sensing target scale and accurately detect ground objects,this paper proposes a target detection and recognition algorithm suitable for remote sensing scenes.The main tasks are as follows:1.Aiming at the problems of data scarcity and multi-scale object interference in target detection in remote sensing scenarios,a small sample remote sensing target detection network based on a two-stage detection model is proposed.Firstly,a novel involution convolution operator is used to construct the detector backbone to improve the feature extraction ability;At the same time,a kind of object pyramid structure is incorporated to enhance the original features to suppress the adverse effects of negative samples,so as to fully mine the feature information of each target scale and help locate semantic information;Finally,the idea of comparative supervision is used to refine the target classification and reduce the false detection rate.Experimental results on publicly available remote sensing datasets show that the network can effectively alleviate overfitting caused by data scarcity and improve detection accuracy when there are only a few labeled samples.Compared with earlier Meta Rcnn and Fs Det networks,the average accuracy has been further improved by 3.8% and 2.5%.2.In order to solve the problem that multiple types of targets in high-resolution remote sensing images have different shapes,small inter class differences,and large intra class differences,resulting in difficulty in fine recognition of detection algorithms.A fine grained recognition algorithm based on fusion of enhanced features is proposed.In the backbone network,the method of combining spatial features with context is used to replace the simple upper and lower layer feature fusion method of the feature pyramid to generate spatial fine grained features;Next,a spatial attention and channel attention enhancement model is used to enhance the attention to spatial details in the neck network,and feature enhancement is performed by grouping on the channel dimension in a constant proportion;Finally,the traditional classification regression dichotomy detection head is improved,and a rotation detection branch is added to incorporate the rotation loss function into the training target,improving the algorithm’s ability to fit targets with different directions in dense scenes.The precision of fine granularity detection on FAIR1 M dataset reached 47.9%,and experimental results show that this method is superior to other existing detection algorithms.3.Aiming at the practical problem of widely distributed and diverse offshore oil production engineering tankers,which makes real-time supervision more difficult,a target detection system for offshore vessels is designed and implemented based on the algorithm trained in this paper.Using a C/S design architecture,the system is mainly divided into a prediction subsystem and a training subsystem.Deconstruct the system from three aspects:the visual layer,the logical layer,and the persistent layer,and describe in detail the roles and internal relationships of each functional module.Track the direction of data flow from the perspective of users,clearly express data requirements and their relationships,and express the logical flow and transformation process of data within the system using a structured system analysis method.Finally,by visualizing the detection of ships on the sea,the effectiveness of this algorithm for remote sensing target detection is more intuitively demonstrated. |
